Checklist
LangGraph human-in-the-loop checklist
Before a LangGraph agent calls a risky tool, add an approval gate that can pause, notify a human, record the decision, and resume only the reviewed action.
1. Classify risky tool calls
- Data writes in staging or production
- Deletes, refunds, purchases, and subscription changes
- External emails, chat messages, or public posts
- Credential, role, permission, and integration changes
- Any tool call involving confidential or regulated data
2. Create a durable approval request
Store the proposed action, target resource, arguments, policy result, agent confidence, reviewer role, timeout, and idempotency key. Do not rely on a transient chat message as the only approval record.
3. Pause before the side effect
The graph should pause before calling the risky tool. If the approval request cannot be created or delivered, fail closed and return a blocked result to the agent.
4. Resume only the approved action
When a reviewer approves, compare the current tool arguments with the reviewed request. If the resource, parameters, policy version, or approval window changed, require a fresh review.
5. Keep an audit trail
- Request ID and idempotency key
- Policy version and policy inputs
- Reviewer identity and decision timestamp
- Final tool execution result
- Expiration, rejection, and cancellation events
Downloadable templates
Ship the approval gate faster.
Get ready-to-edit policies, LangGraph checklist details, OpenClaw/Codex dispatch checks, Feishu/Lark guidance, and production readiness notes.